- Crusoe SVP Kyle Sosnowski says enterprise AI adoption is still early, meaning AI infrastructure demand could keep rising even as companies crack down on token waste
- To control AI costs, enterprises and telcos should start with clear goals and avoid rewarding AI output volume
- Open models could help enterprises cut AI spending for lower-complexity tasks
Crusoe SVP of Cloud Engineering Kyle Sosnowski has a blunt response when someone sends him a seven-page AI-generated document: return to sender.
AI demand is rising, but according to Sosnowski, waste is real. That’s part of the reason costs are skyrocketing, he argued.
While some are worried that the switch from tokenmaxxing to tokenomics will result in a collapse of AI infrastructure demand, Sosnowski told Fierce he doesn’t buy that idea. Outside the confines of edgy, cloud-native Silicon Valley startups, very few companies have fully integrated AI into their operations in a meaningful way, he said.
Indeed, a recent survey from McKinsey found that just 40% of large organizations and 22% of small businesses have move beyond experimenting or piloting AI to scaling it. Among AI tools, chatbots were the most widely scaled (47%) with around 20% stating they were scaling coding and other AI agents.
So, even if companies are reining in token use amid concerns around rising AI costs, there’s plenty of new adoption to offset that pullback, Sosnowski argued. “I think there's still a lot of meat left on the bone,” he said of AI demand.
Dos and don’ts of AI adoption
According to the Crusoe exec, the AI cost problem starts when companies confuse AI usage with AI productivity. He offered a few tips on how to avoid this trap.
First and foremost, he said, enterprises should enter the AI arena “with a goal in mind.” That is, they shouldn’t just be deploying AI to say they’re using it or blindly groping for results. “Make sure you know what you're trying to do first,” he said. That stance will influence everything from token budgets to model selection.
Second, he said companies should be careful not to incentivize volume. “It is taxing to share AI generated documents throughout an organization. It creates a lot of toil on the readers and it spends money needlessly,” Sosnowski explained. “If someone was to share a seven-page AI generated document with me, it is a return to sender moment…I don't care what Claude thinks of the situation. I want to hear what you have to say.”
Maintaining the proper incentives will also help ensure that companies and their employees don’t fall into another trap: outsourcing thinking. AI is great for organizing a person’s thoughts and synthesizing large amounts of information. But using it to offload thinking? That’s downright “dangerous,” he said.
Finally, Sosnowski noted that AI governance is critical for enterprises in an era where coding tools are turning everyone into an engineer. Democratized development is powerful, but without governance it creates data fragmentation and security risk, he argued.
“Building a standardized function inside of your company to be able to ensure they all have the same authentication style and they have the same rigor” is key, he said.
Managing model selection
As telcos begin scaling AI use, Sosnowski said one of the most important cost-control levers will be matching the model to the task. But that’s easier said than done.
“I think it's a problem of education followed by one of tooling,” he said.
Today, he explained, many companies start their AI adoption journey by buying enterprise subscriptions from OpenAI or Anthropic and multiplying the seat cost by the number of employees. But not every use case needs a frontier model. If an employee’s job is largely to summarize five spreadsheets and consolidate the results into one, Sosnowski said, that workload can likely be handled by a cheaper model.
Open models could help fill part of that gap, Sosnowski said, particularly for lower-complexity use cases such as summarization, document organization and internal workflow automation. Frontier models will still have a place for sophisticated reasoning, multi-step workflows and harder technical tasks, he said. But for many enterprise use cases, “good enough” may be good enough.
Read more about enterprise AI and open models here:
Enterprise tokenomics push could unlock new revenue for telcos
Telco AI cost savings could become an opex trap, Bain warns
Enterprises risk AI sticker shock as token costs pile up
Open models are driving AT&T’s AI ‘tokenomics’ strategy
AI enters cost crunch era as hyperscalers zero in on affordability, controls